Skip to content
Meet Fin · A content agent

Fin is the agent behind our blog.

Fin researches, writes, illustrates, publishes, and checks every Field Note. The model handles the thinking. The harness takes care of reliability, access, and recovery. We let Fin deploy articles on his own.

PublishedFin · run replay · 60×
07:25
dispatch to verified live URL
Research
Writing
Publish
Verify
Report
researchGrounded answers with citations saved
researchPrimary pages fetched and saved
writingDraft composed for the fixed reader
checkArticle audit: pass
publishHero rendered, replay hash matches
publishSearch data and metadata checked
publishPost created in the CMS
verifyStored fields read back: match
verifyArticle page 200, title matches
verifyBlog index 200, slug listed
reportPublished. Verified live.
marshal.ing/field-notes/what-is-managed-agent-operations
Live
14
Field Notes Fin has published this month
Every one with a verified live URL
Speed
~7min
From assignment to a checked, live article
Two articles, one run, 12 minutes
Cost
$1.70
AI model cost per article, research through final draft
Before our move to a cheaper model
Hands
0
People touching the work between start and a live URL
We review after, in the report
What an agent is actually made of

The model is three boxes. The job is the other 25.

Most people picture an agent as a model in a box. Fin is a publishing system with a model inside it. The model picks the research, writes the draft, and fixes what the checks flag. Everything else is ordinary code, built so those three decisions can happen safely and be checked afterward.

3/28parts that are the model
ModelCode, checks, tools
Instructions + authority
01Input
Assignment queue we curate
02Context
Instructions, voice, and references
03Access
Only granted tools resolve
04Control
Stop control, checked before every write

Research

Model chooses the sources
05Model
Pick or resume the assignment; decide how much research it needs
06Tool
Search results
07Tool
Answers with citations
08Tool
Primary-page fetch
09Script
Merge research captures

Writing

Model composes and revises
10Model
Write the article
11Check
Mechanical article audit
12Model
One correction turn, never a loop

Publishing

One scripted call
13Check
Confirm the CMS shape
14Script
Render the hero image
15Script
Build and check search data
16Script
Title, description, alt text
17Tool
Upload or reuse the image
18Script
Compose the post
19Tool
Create once, never duplicate

Verification

Same call, then close-out
20Check
Read back what was stored
21Read
Is the article page live?
22Read
Is it on the blog index?
23Decision
Both visible, or stay resumable
24Tool
Mark done and archive
Execution + stored evidence
25Runtime
Sandbox with no network or credentials
26Store
Saved run state, so it resumes
27Runtime
Audited local repairs
28Output
Typed outcome report
Every agent is custom configured.View the engineering schematic
One real run

What we hand Fin, and what comes back.

We wrote the assignment. Fin did everything between it and the report, and nobody touched the work in between.

01We write this
02Fin writes this
03You judge this

AI Assistant vs. AI Agent vs. Managed Service: Who Owns the Job?

Queue
Item #2 of 192
Reader
A founder of a $1M to $10M business. Commercially sharp, not fluent in AI.
Angle
Separate what each option can do from who stays responsible for the work.
Evidence
Saved research and approved material only. No invented experience, quotes, or results.
Published2 of 2 live
Batch run · two articles, in queue order
DispatchedT+00:00
Research calls10
Draft auditsPassed first try
Article 1 liveT+07:25
Article 2 liveT+11:41
Stored fields read backMatch
Article pages200, titles match
Blog index200, both listed
Report deliveredT+11:46
Flagged for you: [verbatim flags from the report]
Glyph-field title card on dark carbon: dense aiAgents texture glowing amber, article title "AI Assistant vs. AI Agent vs. Managed" on staggered dark slabs.

AI Assistant vs. AI Agent vs. Managed Service: Who Owns the Job?

Hero image
Generated from the title in code. Same title, same image, every time.
Byline
Published under our name, like every Field Note.
Read the article
Publishing

The publish that said yes.

What brokeOur CMS accepted every article and returned success. It was also storing each article's search-engine data one level too high, where the site never reads it.
The costEvery Field Note shipped without the data search engines use to understand it. Nothing looked wrong.
Fin nowReads each article back after publishing and checks that every field landed where the site actually reads it.
Verification

The check that trusted bad data.

What brokeArticles in multi-word categories were saved with a value the CMS didn't recognize. Our first fix learned the valid values from past articles. They all shared the same wrong value, so the check approved the mistake.
The costArticles silently lost their category, and the safeguard vouched for the error.
Fin nowChecks against a list we maintain alongside the live schema, never against the data it is checking.
Verification

The page that wasn't there yet.

What brokeAn article can be saved before the site serves it. If the first visit loses a race with the site's cache, readers can get "not found" for up to five minutes.
The costA "published" article nobody could open, and nobody told.
Fin nowChecks the live page and the blog index before calling the job done. If it can't confirm both, it says so and finishes on the next attempt without duplicating anything.
Cost

The $9.10 article.

What brokeOur first end-to-end run took 101 model steps. Only 22 were writing. The rest were the agent coordinating itself, paying AI prices for clerical work.
The cost$9.10 and 35 minutes for a single article.
Fin nowUses the model only where judgment is needed. A later run published two articles in 62 steps, 12 minutes, and $3.40 combined. [current cost per article]

We found and fixed each of these ourselves. The business an agent works for should never have to. That's what managed means.

Evaluate the writing

Fin delivers finished work.

Managed Agent Operations

Fin saves us an entire workday, each week.

Nothing chews up more time than content creation. We built Fin to solve this problem once and for all.

Everything is pre-approved before it ships
Glyph-field title card on dark carbon: dense workflows texture glowing purple, article title "Automate Legal Intake: From Inquiry to" on staggered dark slabs.
Workflows

Automate Legal Intake: From Inquiry to Attorney Review

Glyph-field title card on dark carbon: dense aiAgents texture glowing purple, article title "How to Build an Honest AI Business Case" on staggered dark slabs.
AI Agents

How to Build an Honest AI Business Case

Glyph-field title card on dark carbon: dense aiAgents texture glowing cyan, article title "Done-for-You AI: What Should Be Included?" on staggered dark slabs.
AI Agents

Done-for-You AI: What Should Be Included?

If an agent ran your blog

Same system. Your topics, your voice, your site.

You provideThe topics you want covered, who you're writing for, how you sound, and anything that's off limits. We turn that into the assignment queue.
It publishes toYour existing site: WordPress, Webflow, Shopify, HubSpot, etc. You authorize access. The agent never sees your password.
You getA live article and a report on every run, with anything that needs your attention at the top.
We handleThe build, the checks, and the fixes when your site, your CMS, or an AI provider changes something.
01
Assign
02
Research
03
Draft
04
Publish
05
Verify
06
Report
Where you come in: after the report. You read what went live and tell us what to change.

What checks can't do. Automated checks catch structure, formatting, and rule problems. They can't guarantee an article is accurate or worth reading. Editorial judgment stays with you, which is why many owners start with sign-off.

Before you outsource your blog to a machine

Questions owners ask

Will AI-written articles hurt my search rankings?

Search engines judge whether content is useful, not which tool produced it. Fin writes to a specific reader, uses saved research, and isn't allowed to invent experience or results. Every article also ships with the on-page information search engines use, and Fin confirms that information actually landed on the page.

Can it write in my voice?

Yes. We build the writing instructions from your existing material and your standards. We read what it publishes and tune its skills accordingly.

Who is accountable if it publishes something wrong?

We are accountable for every agent. When it comes to content creation, you decide what the agent may publish and its level of autonomy. Automated checks catch structural and formatting problems. And if you feel more comfortable with a person between the draft and your site, choose sign-off and Marshal sends you the article for review.

Can I approve articles before they go live?

Yes. Sign-off is a setting we engage before the agent starts. You can move to direct publishing later, or back. It's entirely up to you.

What does it cost?

Plans start at $249 a month per workload, with no setup or integration fees. What you're buying is fully completed work and the ongoing management of your agent.

Let's get back tothe work of humans⁠1

Machine-like is the services layer that bridges frontier AI and business operations.